Targeted training manipulation can erase a single language from an LLM while leaving a related one intact

Dan Balsam describes a technique that, if it holds up at scale, reframes one of the more vexing problems in language model control. Using parameter decomposition and targeted training manipulation, he says, a researcher can cause a model to forget a single language while leaving a closely related one intact. The example he gives is specific: German goes, Dutch stays.

That specificity matters. German and Dutch are closely related, sharing vocabulary, grammatical structures, and script. Anyone familiar with multilingual models would expect the two to be entangled in the parameter space in ways that make selective forgetting look close to impossible. If you pull hard enough on German, the intuition runs, you should dislodge Dutch along with it. Balsam’s claim is that the technique threads that needle.

The mechanism he names is parameter decomposition. The idea is not that training is simply run in reverse or that a blunt fine-tuning pass degrades one language. Instead, the parameters encoding a target language are identified and manipulated with enough precision that neighboring linguistic knowledge survives. Whether that precision holds across model sizes, training regimes, and language pairs beyond the German-Dutch case, the claim does not say. But the framing is methodological, not anecdotal: Balsam presents this as something a researcher can do, not something that happened once by accident.

We can get LM to forget a single language. We can get LM to forget German and not forget Dutch. Dan Balsam

The reason that framing carries weight is that machine unlearning, as the field sometimes calls it, has historically been a problem of collateral damage. Techniques that scrub one kind of knowledge from a model tend to degrade adjacent capabilities in ways that are hard to predict and harder to audit. A method that can isolate a single language without disturbing a linguistically proximate neighbor would represent a meaningful step toward surgical model editing, the ability to remove or modify a specific capability without touching the broader architecture of what the model knows and can do.

The practical applications are not hard to sketch. Regulatory requirements in some jurisdictions restrict the languages a deployed model may use. Operators who need a model that will not respond in German, whether for compliance, licensing, or market-segmentation reasons, currently have limited options short of output filtering, which can be circumvented, or training a separate model from scratch, which is expensive. A parameter-level approach that reliably removes the capability at the source would change that calculus.

There is also a darker mirror to the application. A technique that can remove a language selectively can, in principle, be applied to other capabilities that occupy a similarly bounded region of the parameter space. The research question Balsam’s claim opens is whether language turns out to be unusually well-localized, making it a special case, or whether the same decomposition logic can be extended to factual domains, reasoning styles, or behavioral tendencies. That question is not answered by the claim. It is the question the claim makes newly pressing.

What Balsam describes is a proof of concept with a clean, falsifiable structure. Either a model trained with this method can produce fluent Dutch after losing German, or it cannot. Either the forgetting is durable across evaluation sets, or it degrades. The claim invites exactly that kind of testing, and the fact that it is built around a hard pair of related languages, rather than an easy pair of distant ones, suggests the technique was designed to face the difficult version of the problem from the start.